RFR: 8294186: AArch64: VectorMaskToLong failed on SVE2 machine with -XX:UseSVE=1

[Andrew Haley]

On Tue, 27 Sep 2022 09:48:24 GMT, Eric Liu <eliu at openjdk.org> wrote:

> C2_MacroAssembler::sve_vmask_tolong would fail on BITPERM supported SVE2 machine with "-XX:UseSVE=1".
> 
> `BITPERM` is an optional feature in SVE2. With this feature, VectorMaskToLong has a more efficent implementation. For other cases, it should generate SVE1 code.
> 
> [TEST]
> jdk/incubator/vector, hotspot/compiler/vectorapi passed on BITPERM supported SVE2 machine, with option -XX:UseSVE=(0, 1, 2).

src/hotspot/cpu/aarch64/c2_MacroAssembler_aarch64.cpp line 1001:

> 999:     // Given by the vector with value 0x00 or 0x01 in each byte, the basic idea
> 1000:     // is to compress each significant bit of the byte in a cross-lane way. Due
> 1001:     // to the lack of cross-lane bit-compress instruction, here we use BEXT

Suggestion:

    // Given a vector with the value 0x00 or 0x01 in each byte, the basic idea
    // is to compress each significant bit of the byte in a cross-lane way. Due
    // to the lack of a cross-lane bit-compress instruction, we use BEXT
    // (bit-compress in each lane) with the biggest lane size (T = D) then
    // concatenate the results.```

-------------

PR: https://git.openjdk.org/jdk/pull/10443